Deep Learning-Driven Dynamic Network DEA for Cross-Industry ESG Resilience: Heterogeneous Threshold Identification and Carbon Policy Simulation

Balancing production resilience with environmental, social and governance (ESG) performance is difficult when disruptions, policy constraints and stakeholder expectations interact over time. This paper proposes a six-node dynamic network data envelopment analysis architecture whose admissible weight intervals are adjusted by an LSTM learner using IoT-derived shock states. The application covers 96 aggregate monthly periods from 2018 to 2025 across three countries and four technology-based manufacturing groups. Figure-grounded diagnostics indicate mean resilience scores of 0.7016 for the basic IoT-DEA benchmark and 0.7134 for the complete model (paired difference = 0.0118; 12-month moving-block bootstrap 95% CI = 0.0015–0.0245; p = 0.001), while mean uncertainty falls from 0.0536 to 0.0205, a 61.8% reduction. Threshold sensitivity shows that governance-delay and skill boundaries vary by industry and carbon-constraint severity rather than constituting universal standards. In the calibrated policy simulation, the combined carbon-tax and green-subsidy path reaches approximately 0.90 by month 96, compared with 0.79 under no policy; this contrast is interpreted as a scenario result, not a firm-level causal treatment effect. NSGA-III and SHAP then connect the measured constraints to Pareto-efficient portfolios and sector-specific managerial priorities. The framework’s main supported contribution is more stable, temporally explicit ESG-resilience diagnosis with transparent limits on causal and cross-sectional inference.